LoFTR vs EfficientLoFTR predicitons
williamhoole opened this issue · comments
Hi,
I have noticed that LoFTRs predictions are mostly on the defined grid that is used for coarse matching. I have noticed that in EfficientLoFTR this is not necessarily the case and that it has more refined predictions. Why is this and is the data and supervision done differently during training ?
(note: i have changed the image size to 256x256)
Original LoFTR:
Efficient LofTR:
I will close this task, I have seen that the fine window size in the Original LoFTR is 5x5 and in the EfficientLoFTR it is 8x8. Im guessing this in general means that the efficient LoFTR is capable of more accurate matches. Here is an example where i Rotated the images by 5°. You can see that both models are able to predict at fine level outside of the coarse level grid.
Original LoFTR with 5° rotation:
Efficient LoFTR with 5° rotation:
